{"id":6728,"date":"2025-05-12T03:27:00","date_gmt":"2025-05-11T18:27:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=6728"},"modified":"2025-05-11T07:29:58","modified_gmt":"2025-05-10T22:29:58","slug":"teaching-models-to-understand-but-not-generate-high-risk-data","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=6728","title":{"rendered":"Teaching Models to Understand (but not Generate) High-risk Data"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Teaching Models to Understand (but not Generate) High-risk Data\u00a0<\/strong>[38.3]<br>SLUNG(Selective Loss to Understand but not Generate)\u3092\u7d39\u4ecb\u3059\u308b\u3002 SLUNG\u306f\u3001\u30e2\u30c7\u30eb\u304c\u9ad8\u30ea\u30b9\u30af\u30c7\u30fc\u30bf\u3092\u751f\u6210\u305b\u305a\u306b\u7406\u89e3\u3059\u308b\u3053\u3068\u3092\u5b66\u3076\u305f\u3081\u306e\u4e8b\u524d\u5b66\u7fd2\u30d1\u30e9\u30c0\u30a4\u30e0\u3067\u3042\u308b\u3002 SLUNG\u306f\u3001\u751f\u6210\u3092\u5897\u5927\u3055\u305b\u308b\u3053\u3068\u306a\u304f\u3001\u30e2\u30c7\u30eb\u306b\u3088\u308b\u9ad8\u30ea\u30b9\u30af\u30c7\u30fc\u30bf\u306e\u7406\u89e3\u3092\u4e00\u8cab\u3057\u3066\u6539\u5584\u3059\u308b\u3053\u3068\u3092\u793a\u3059\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2505.03052v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2505.03052v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Mon, 05 May 2025 22:24:06 GMT)<\/li>\n\n\n\n<li>\u300cThis work introduces SLUNG, a pre-training paradigm that enables language models to learn from high-risk data without being trained to generate it. By selectively adjusting the training objective at the token level based on risk, SLUNG decouples a model\u2019s ability to understand from its ability to generate, allowing models to condition on high-risk inputs while learning from adjacent low-risk tokens.\u300d\u3068\u3044\u3046\u624b\u6cd5\u306e\u63d0\u6848\u3002\u53e3\u5916\u3059\u308b\u3053\u3068\u306f\u3067\u304d\u306a\u3044\u304c\u5b66\u3076\u5fc5\u8981\u304c\u3042\u308b\u3082\u306e\u3001\u3068\u3044\u3046\u306e\u306f\u73fe\u5b9f\u7684\u306b\u591a\u3044\u308f\u3051\u3067\u3053\u306e\u3088\u3046\u306a\u624b\u6cd5\u306f\u975e\u5e38\u306b\u9762\u767d\u3044\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[223,352],"class_list":["post-6728","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-llm","tag-security"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6728","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6728"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6728\/revisions"}],"predecessor-version":[{"id":6729,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6728\/revisions\/6729"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6728"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6728"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6728"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}